Liquid AI has released LFM2.5-VL-3B for image processing and function calling
The model LFM2.5-VL-3B is available on Hugging Face and in Playground. It supports function calling from text and image inputs. According to Liquid AI, it improves screen understanding, object localization and handling multiple images.
Liquid AI has released LFM2.5-VL-3B, a model for processing images and text with 3.1 billion parameters. It is available on Hugging Face and in Playground. The company also provides documentation for running it locally and fine-tuning it.
Function-calling support is a new addition to the company’s vision model lineup, working with both text-only input and combined image and text input. According to the company, the model better understands mobile, web and desktop application interfaces, more accurately links text instructions to object locations in images, and handles multiple images simultaneously more effectively. The model responds directly; the company is targeting applications that require low latency and on-device use.
According to measurements by the company, the model achieves an average score of 80.7 on ScreenSpot-v2. Compared with the previous model LFM2-VL-3B, the score on ToolSandbox increased from 26.4 to 59.5, and on BFCL v4 from 20.5 to 32.5. Accuracy in locating objects based on text instructions on RefCOCO rose from 57.1 to 87.9. These results were published by the manufacturer.
Why it matters
Support for image inputs and function calling makes it possible to build applications that select subsequent operations based on screen content. The availability of the model and documentation for running it locally gives developers a way to test such a solution on their own tasks; the published scores alone do not confirm reliability in a specific deployment.
Release card
LFM2.5-VL-3B
Liquid AI
- Specifications
- The model has 3.1 billion parameters.
- Inputs
- The model accepts text and image inputs, including multiple images, and responds with text. It also supports function calling.
- Availability
- The model is available on Hugging Face and in Playground. Liquid AI provides documentation for running it locally and fine-tuning it.
- ScreenSpot-v2 80.7 According to measurements by Liquid AI, this is the average score for identifying elements on screens in mobile, web and desktop applications.
- ToolSandbox 59.5 In a tool-use evaluation conducted by Liquid AI, the model scored 59.5 compared with 26.4 for the previous model LFM2-VL-3B.
- BFCL V4 32.5 In a function-calling evaluation conducted by Liquid AI, the model scored 32.5 compared with 20.5 for the previous model LFM2-VL-3B.
- RefCOCO-avg 87.9 For locating objects based on text instructions, Liquid AI reports a precision@1 of 87.9 compared with 57.1 for the previous model LFM2-VL-3B.
- BLINK 61.5 In a multi-image understanding evaluation conducted by Liquid AI, the model scored 61.5 compared with 50.2 for the previous model LFM2-VL-3B.
- Recognizing elements in mobile, web and desktop application interfaces.
- Calling functions based on text input or combined image and text input.
- Locating objects in images based on text instructions.
Liquid AI compares the model with the previous model LFM2-VL-3B and reports improvements in screen understanding, object localization, function calling and handling multiple images. The model responds directly without a separate reasoning phase, and the company is targeting low-latency applications and on-device use.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
Two audiences, two different impacts
What this means
For individuals
Individuals can use Playground to try recognizing screenshot content and comparing multiple images without setting up the model locally.
For a business
Development teams gain a model for prototypes that connect screen processing with function calling. The documentation also covers running the model locally and fine-tuning it for business tasks.
DevelopmentCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.